Opus 5 vs GPT6 Astra: Claude Codes an SMB Clone 9 Minutes Faster and Sticks to the Prompt
devino21 · reddit · 2026-09-20
Using the same game-clone prompt (an SMB-style game) built for testing local LLMs, the author compared Claude Code + Opus 5 High against Codex + GPT6 Astra High:
- Claude Code + Opus 5: 16:56, 33% of the 5-hour quota, output closely matched the original game (named Fungus Frontier).
- Codex + GPT6 Astra: 25:57, 55% of quota, but "did what it wanted" — completely different color theme, shrunk to 640x480, renamed to its own theme (Embertrail).
The author was leaning toward Astra and nearly cancelled, but this test made him reconsider. A same-task head-to-head highlighting a clear instruction-following gap.
More from coding & agent
- Devs debate running stateful AI agent runtimes on Cloudflare Workers and other edge runtimes — merlinofthewater · 2026-09-20
- Jev's instant compression scores each tool call to trim agent context without summarization — FinanceYF5 · 2026-09-20
- Stop using LLMs for ticket triage: Jev turns unstructured input into an executable smart if — sven_ai · 2026-09-20
- Jeff Dean's 1-Hour AI Engineering Lecture: From LLM Basics to Agent Graphs — irinarish · 2026-09-20
- Multi-agent systems work best with clear roles, not more agents — _jaydeepkarale · 2026-09-20
- Jev founder: all AI models are built for human-in-the-loop, not true automation — hardimanjames · 2026-09-20